Comparing machine learning techniques to segmentize and classify tongue regions for traditional and complementary medicine (TCM) diagnosis

Bong, Min Xuan (2025) Comparing machine learning techniques to segmentize and classify tongue regions for traditional and complementary medicine (TCM) diagnosis. Final Year Project, UTAR.

PDF Download (2360Kb)

Abstract

This project investigates the application of machine learning and deep learning techniques for automated tongue diagnosis in the context of Traditional Chinese Medicine (TCM). Tongue diagnosis, a long-established diagnostic method in TCM, is often limited by subjectivity and inconsistency. To address this, the study develops a systematic pipeline that integrates segmentation and classification models, enabling more objective, accurate, and reproducible analysis of tongue images. Three datasets—binary (stained vs. non-stained moss), four-class (color variations), and five-class (coating categories)—were utilized to evaluate performance under varying levels of complexity. Segmentation was performed using both classical methods (SVM) and a deep learning approach (DuckNet), with DuckNet providing superior accuracy and robustness. Classification was carried out through an evolutionary series of architectures, beginning with AdderNet and progressing through ResNet20, HybridNet, and an Improved HybridNet. Experimental results demonstrated that while AdderNet achieved the highest accuracy in complex multi-class scenarios, it suffered from excessive computational cost and scalability limitations. The Improved HybridNet consistently offered the best trade-off between performance and efficiency, delivering strong accuracy with reduced parameters, training time, and model size. Overall, the project highlights the potential of artificial intelligence to modernize tongue diagnosis by providing standardized, efficient, and clinically relevant computational tools. The findings establish a foundation for future integration of AI-driven diagnostic support systems into healthcare practice.

Actions (login required)